Parallel beamlet dose calculation via beamlet contexts in a distributed multi‐GPU framework

• DOI: 10.1002/mp.13651
• Publication Date: 2019-06-11T03:08:08Z
• Authors: Ryan Neph; Cheng Ouyang; John Neylon; Youming Yang; Ke Sheng
• Keywords: Radiotherapy Planning; Oncology and Carcinogenesis; GPU; Biomedical Engineering; Bioengineering; dose calculation; Radiation Dosage; Phantoms; Imaging; distributed computing; 03 medical and health sciences; Computer-Assisted; 0302 clinical medicine; Intensity-Modulated; Computer Graphics; convolution; Radiometry; Medical and biological physics; Radiotherapy; Phantoms, Imaging; Radiotherapy Planning, Computer-Assisted; Radiotherapy Dosage; Medical and Biological Physics; Other Physical Sciences; Nuclear Medicine & Medical Imaging; Physical Sciences; Radiotherapy, Intensity-Modulated; beamlet; Biomedical engineering; Monte Carlo Method

Dose calculation is one of the most computationally intensive, yet essential tasks in treatment planning process. With recent interest automatic beam orientation and arc trajectory optimization techniques, there a great need for more efficient model-based dose algorithms that can accommodate hundreds to thousands candidates at once. Foundational work has shown translation graphical processing units (GPUs), lending remarkable gains efficiency. But these methods provide parallelization only single beamlet, serializing multiple beamlets under-utilizing potential modern GPUs. In this paper, authors propose framework enabling parallel computation many beamlet doses using novel context transformation further embed approach scalable network multi-GPU computational nodes.The proposed context-based separates beamlet-local density TERMA into distinct contexts independently sufficient data calculation. Beamlet are arranged composite array with dosimetric isolation, subjected GPU collapsed-cone convolution superposition procedure, producing set beamlet-specific distributions pass. from each converted sparse representation storage retrieval during plan optimization. The radius new parameter permitting flexibility between speed fidelity A distributed manager-worker architecture constructed around supporting an arbitrary number worker nodes resident Phantom experiments were executed verify accuracy compared Monte Carlo reference CPU-CCCS implementation broad beams composed by addition beamlets. representative 4π sets was calculated lung prostate cases compare its efficiency existing beamlet-sequential GPU-CCCS implementation. Code profiling also performed evaluate scalability across networked GPUs.The method displays <1.35% 2.35% average error serialized algorithm simulation PDDs water slab phantoms, respectively. demonstrates substantial speedup up two orders magnitude over tested configurations. near linear scaling compute GPUs employed, indicating it flexible enough meet performance requirements users simply increasing hardware utilization.The expectation beamlet-based methods. This been successful accelerating process very large-scale problems - such as IMRT VMAT selection, clinically feasible timeframes. makes strong candidate use variety other clinical workflows.